Hi Derek,
I haven't tried with lam-mpi, but with intel mpi there are problems
if you link with a blacs compiled for mpi 1.2 while using mpi 2
(lam-mpi implements almost all of mpi 2). The problem is in one
of the include files of the c interface of blacs, so it can be really
difficult to detect while using a fortran only application. In order
to link with mkl I'm using the following:
GUIDE=/opt/intel/cmkl/8.1.1/lib/em64t/libguide.a
SCALAPACK=/opt/intel/cmkl/8.1.1/lib/em64t/libmkl_scalapack.a
#Pay attention to the intelmpi20 suffix of blacs:
BLACS=/opt/intel/cmkl/8.1.1/lib/em64t/libmkl_blacs_intelmpi20.a
VML=/opt/intel/fce/9.1/lib/libsvml.a
LAPACK=/opt/intel/cmkl/8.1.1/lib/em64t/libmkl_lapack.a
BLAS=/opt/intel/cmkl/8.1.1/lib/em64t/libmkl_em64t.a
LIBS=$(SCALAPACK) $(BLACS) $(LAPACK) $(BLAS) $(GUIDE) $(VML) -lpthread
Also if nuotot is only 24 then the system is really small, maybe
there are no orbitals in a given node and it produces the NaN.
Good luck, and please post your findings!
Eduardo
On oct 5, 2006, at 5:32 PM, Derek A. Stewart wrote:
Hi everyone,
Just a quick follow up on the NaN errors I have been running
into with
Siesta 2.0. It looks like the error occurs in the cdiag subroutine
(cdiag.F) when pzheevd is called to solve the standard eigenvalue
problem. I have tried changing the parameters of the run (i.e.
DivideConquer, Set2D), but similar problems occur.
This means it is probably a problem with the scalapack library. Has
anyone successfully compiled the parallel version of Siesta with
lam-mpi
7.1.1 with intel compilers version 8 or 9? If they have, I would be
interested in comparing your Bmake, SLmake input files for BLACS and
SCALAPACK with what I am using.
Thanks,
Derek
On 10/4/06, Derek A. Stewart <[EMAIL PROTECTED]> wrote:
Hi all,
I have been trying to get the parallel version of Siesta 2.0
to run
with the intel compilers (version 8.1, MKL math libraries) and
lam-mpi
(7.1.1). I have no problems compiling the parallel version of
siesta.
However, when I run a calculation, I run into a problem which I have
traced to the diagk.F file. Evidently the psi matrix
psi(2,nuotot,nuo) which provides auxiliary space for the
eigenvectors
has NaN entries for nuo=1,2. The other entries (nuo=3-24) are fine.
This can probably be traced to the cdiag subroutine called in this
file and I am working on this.
Has anyone run into this problem before? For my compilation, it
appears
to occur for general input files. The run has no problems for the
serial
case.
Thanks,
Derek
###########################
Derek Stewart, Ph. D.
250 Duffield Hall
Cornell Nanoscale Facility
stewart (at) cnf.cornell.edu
(607) 255-2856
